train/validate minc-2500 dataset in chainer

MINC-2500 dataset: download

Material Recognition in the Wild with the Materials in Context Database
Sean Bell and Paul Upchurch and Noah Snavely and Kavita Bala,
Computer Vision and Pattern Recognition (CVPR), 2015.

• chainer 1.15+
• numpy
• scikit-learn
• Pillow
• matplotlib
• tqdm
• dominate

• put ''minc-2500'' directory in ''chainer-minc-2500'' directory

• make image-label dataset before training

  python make_dataset.py ./minc-2500 shuffled_labels --shuffle
  python make_all_list.py
  

• if using fine-tuning, download mean file and imagenet-pretrained model

  python download_mean_file.py
  

  python download_model.py
  

• available pre-trained architechures

  • AlexNet
  • GoogLeNet(v1)
  • VGG (16 layers and 19 layers)
  • NIN (Network-in-Network)
  • SqueezeNet (ver. 1.0 and ver. 1.1)

  see Caffe Model Zoo for more details.

example: train(fine-tune) GoogLeNet with GPU

python train_minc2500.py ./minc-2500/shuffled_labels/train1.txt \
./minc-2500/shuffled_labels/validate1.txt -a googlenet --finetune -E 10 -R ./minc-2500 -g 0

python randomsearch.py ./minc-2500/shuffled_labels/train1.txt \
./minc-2500/shuffled_labels/validate1.txt --finetune -E 1 -R ./minc-2500 -g 0 -t 20

python gridsearch.py ./minc-2500/shuffled_labels/train1.txt \
./minc-2500/shuffled_labels/validate1.txt --finetune -E 1 -R ./minc-2500 -g 0

example: visualize "conv1" layer of imagenet-pretrained AlexNet (bvlc_alexnet.caffemodel) using all minc-2500 images

1. extract filter output
   このプログラムは、指定した層の各フィルタ（全結合層の場合は各ニューロン）を刺激する上位9枚(-t 9)の画像を調べます。

python extract_features.py ./minc-2500/all_list.txt -a alex --finetune -b 50 -g 0 -R ./minc-2500 \
-m ilsvrc_2012_mean.npy -l conv1 -t 9

This program generates "./result/alex/extract/features/top_conv1.txt".

2. acquire most activated image patch
   このプログラムは、手順1の結果をもとに、活性化パッチ画像を取得します。

python acquire_patches.py ./minc-2500/all_list.txt -a alex --finetune -b 50 -g 0 -R ./minc-2500 \
-m ilsvrc_2012_mean.npy -l conv1

Please execute procedure 1 before executing this procedure.
This program generates images (./result/alex/extract/conv1/*.png).
This program also generates "./result/alex/extract/features/maxbounds_conv1.txt" and "./result/alex/extract/features/maxloc_conv1.txt".

3. acquire deconv image
   このプログラムは、手順1の結果をもとに、Zeiler & Fergusの手法を用いて可視化画像を取得します。
   （このプログラムを実行する前に、手順2を実行する必要はありません）

python acquire_patches.py ./minc-2500/all_list.txt -a alex --finetune -b 50 -g 0 -R ./minc-2500 \
-m ilsvrc_2012_mean.npy -l conv1

Please execute procedure 1 before executing this procedure (You don't have to execute procedure 2).
This program generates images (./result/alex/extract/deconv_conv1/*.png).

You can concatenate images using concat_image.py
先の例のようにフィルタ毎に上位9枚の画像を出力した時、画像はバラバラで出力されますが、 下のようにすると、9枚の画像が3x3の正方形状に連結されます。

python concat_image.py -R ./result/alex/extract/conv1 -c 3